Introduction

Chemical probes — potent, selective, cell-active small molecules targeting specific proteins — constitute some of the most impactful research tools in the life sciences arsenal, as evidenced by citations and impact on drug discovery1,2. The broader availability of chemical probes for all human proteins would greatly advance our understanding of the human proteome, as well as help prioritize potential new drug targets. In 2009, the Structural Genomics Consortium (SGC) launched a programme to assemble and invent chemical probes for human proteins related to cell signalling, protein homeostasis and epigenetics. The programme successfully developed and collected new chemical probes for over 200 unique proteins from the academic and industrial communities. The impact of these 200 chemical probes has been profound: more than 60,000 samples have been distributed to scientists around the world, they have collectively been cited at least 13,000 times as assessed by searching for the name of the probe in Google Scholar, and the discoveries they have enabled are being tested in more than 85 clinical trials.

The obligatory first step in creating a chemical probe for a new protein (or a proximity pharmacology tool such as proteolysis-targeting chimeras (PROTACs))3 is to identify a validated, chemically tractable hit. For proteins that belong to precedented classes of drug targets, hits can often be identified quite readily, either by screening focused chemical libraries that are enriched in experimentally verified structural classes4 or by making computational predictions based on pre-existing experimental data5,6,7,8,9,10. By contrast, for lesser-studied proteins, hit-finding is more challenging and is often rate determining. Currently, hit finding is almost always initiated with an experimental screen of large and diverse chemical libraries followed by time- and cost-intensive hit verification and optimization. Although the available experimental hit-finding approaches have expanded greatly over the past 20 years, there has not been a dramatic improvement in their overall success rates or cost effectiveness11,12,13. This situation underscores the need for a radically different approach in the context of the Target 2035 initiative14.

Computational methods, particularly machine learning (ML) and artificial intelligence (AI) strategies have the most potential to develop cost-effective hit-finding methods for unprecedented targets15. However, the development of hit-finding algorithms is currently limited by the lack of suitable protein–ligand datasets in the public domain16,17: the existing chemical bioactivity datasets are either fragmented across databases such as ChEMBL and PubChem, or are not available to the public18, most have been compiled from non-standardized experimental protocols that introduce noise into training data19, the datasets are not always prepared for ML/AI analysis, and most lack high-quality data on inactive compounds20.

With data paucity identified as the greatest hurdle to the development of hit-finding algorithms, our SGC/Target 2035 working group decided that the next phase of the Target 2035 initiative (2025–2030) should organize a programme that (1) systematically generates large experimental protein–small–molecule binding datasets and provides open access to the well-annotated data, and (2) works with the community to train, develop, refine, test and benchmark hit-finding algorithms, to start.

A scientific and operational plan for the initiative, including target selection, data generation and dissemination, benchmarking of ML/AI predictions, success criteria, governance, and funding was discussed in a face-to-face meeting in Frankfurt, Germany, in the autumn of 2023, and in London, UK, in the autumn of 2024. In this Roadmap, we consolidate the outputs from these meetings into an ambitious yet tractable roadmap to provide sufficient experimentally derived data to transform hit finding into a computational endeavour. We also highlight how the Target 2035 open science initiative is structured to provide ample mechanisms for the greater experimental and computational academic and industry communities to contribute and benefit.

In our approach, there are conceptual parallels between our proposed approach and the development of the AlphaFold programs for protein structure prediction. The successful application of ML to protein structure prediction was empowered by massive open data generation by the structural biology and genomics communities, longstanding stewardship of the data by the Protein Data Bank and GenBank teams21 and an engaged structure prediction community, whose algorithms were benchmarked by the CASP team (Critical Assessment of Protein Structure Prediction)22 through open challenge competitions. This analogy has limits though. The immense space of intramolecular interactions that is afforded by 20 amino acids and defines the protein-folding paradigm is relatively constrained when compared with the diversity of possible interactions between proteins and ~1060 drug-like molecules. Clearly, novel ML strategies will be necessary for a breakthrough in AI-driven drug design, and it is not possible a priori to predict the size and diversity of the protein–ligand datasets that will be required to enable such a breakthrough, or even if it will be possible in the near term. With this caveat, it is nevertheless apparent that high-quality, large-size protein–ligand datasets will be foundational to solving the problem.

Overview of project workflow

This 5-year project will generate high-quality, open datasets that include binding data for millions to billions of small molecules to more than 2,000 diverse proteins. The data will include results from testing both experimentally derived and computationally predicted hit candidates using orthogonal biophysical and functional assays.

The project workflow is outlined in (Fig. 1) and is described in more detail below. In brief, the project will

  1. (1)

    Generate purified proteins both within the project and by inviting community members to contribute purified proteins. All purified proteins would be subject to strict quality control.

  2. (2)

    Generate binding data using affinity-selection mass spectrometry (AS–MS) and DNA-encoded chemical library (DEL) screening, each of which measures the binding of small molecules to purified proteins directly. AS–MS and DEL screening are also performed in a standardized way, and the outputs have associated quality metrics. Candidate small-molecule binders will be tested in secondary screens using orthogonal, high-quality biophysical assays.

  3. (3)

    Make annotated primary screening data openly available in an ML/AI-ready format via a project database called AIRCHECK (Artificial Intelligence-Ready CHEmiCal Knowledge base; https://aircheck.ai/).

  4. (4)

    Challenge the ML/AI and computational chemistry communities to make predictions based on the data and organize a series of benchmarking competitions to help advance the methods.

  5. (5)

    Experimentally test community predictions using biophysical methods.

  6. (6)

    Share assay data from predicted binders via AIRCHECK.

  7. (7)

    Share reagents, protocols, binders and data without restrictions on use.

Fig. 1: Data generation pipeline.
figure 1

The workflow for generating data and binders. (1) Purified proteins are produced in experimental hubs, by partners or by the community. (2) Proteins are screened in project screening laboratories, and data are experimentally annotated in partner laboratories or experimental hubs. (3) Quality-controlled (QC) screening data are deposited into the AIRCHECK database. (4) Computational experts in the project and the community build machine learning (ML)/artificial intelligence (AI) models and make predictions about new or improved binders. (5) Predicted compounds are procured and tested in experimental hubs. (6) The QC’d assay data, including hits and binding data, are deposited into the Artificial Intelligence-Ready CHEmiCal Knowledge base (AIRCHECK). (7) Hits and data are released to the community, freely available for further research and development.

Full size image

In addition, for as many confirmed binders as feasible, co-crystallization with the cognate target would be attempted, and structure–activity relationships explored by testing structural analogues of the confirmed binders, either purchased from vendors or synthesized by collaborating chemists. Ideally, all binders would be tested in functional assays when available.

The project will have two important outcomes. First, it will generate new small-molecule binders for prioritized proteins. Second, it will create a comprehensive, well-annotated dataset to advance computational methods. This second outcome will be achieved by prioritizing data quality, data consistency and data access, and designing the experimental workflow and outputs in partnership with data scientists23.

Access to diverse high-quality proteins

To generate protein–small-molecule binding datasets of sufficient size and diversity, it will be critical to access and prosecute a structurally diverse set of purified, homogeneous and stable proteins. Given that it is not possible a priori to estimate the number of datasets that will be required to enable computational methods, for planning purposes, we have arbitrarily set a goal to screen a minimum of 2,000 different proteins over 5 years. For perspective, and to attest to the feasibility of the project, this is approximately the number of unique proteins purified within the SGC in the 5-year period between 2007 and 2012.

The selection of which proteins to screen will be guided by the requirement to maximize the structural and functional diversity of the protein targets, as well as by the desire of funders and participants to identify hits for protein targets of their immediate scientific interest. Initially, experimental tractability will be prioritized to establish and optimize project platforms, logistics, data workflows and procedures. Tractable targets constitute those that can be readily purified in sufficient quantities, are known to have suitable biophysical properties, and for which orthogonal assays are either already available or can be readily developed (Fig. 2). The SGC and the wider protein and structural biology communities have successfully produced over a thousand human proteins (or domains thereof) in the past that meet these criteria, and these proteins should be easily and rapidly accessible or able to be repurified. A graphical representation of ~400 proteins already purified at the SGC is included in Supplementary Fig. 1 and a snapshot of the protein database in Supplementary Table 1. As the project progresses, the number of never-before purified proteins and proteins that are more technically challenging to produce will be increased.

Fig. 2: Target selection and prioritization.
figure 2

The protein pipeline will comprise proteins that have been produced previously by project participants, new proteins that are nominated by the funders, and proteins contributed by experts in the wider community. Among the proteins, the project will create a Target List that integrates structural and ligand-binding pocket diversity, funder interests and scientific priorities. Proteins produced previously will be prioritized at the outset to focus on logistics and to generate data, and never previously produced proteins will be added as the project progresses. PDB, Protein Data Bank; SGC, Structural Genomics Consortium.

Full size image

Protein production

To ensure protein quality and consistency, protein quality criteria have been established and implemented (an exemplar is shown in Supplementary Fig. 2) and the majority of the proteins will be produced in a handful of geographically distributed protein purification hubs that share methodologies. These hubs will probably be organized around protein families and/or scientific themes. To attract a wider diversity of protein targets, experts in the community would be invited to contribute purified proteins that meet the diversity and quality criteria. The incentives for community members to contribute proteins will be to access high-quality chemical screening capabilities and to be able to pursue any small-molecule hits identified in the screens, without precondition (https://public.thesgc.org/protein_registry/protein_intake.php).

Protein–ligand open data generation

All purified proteins that pass quality control will be screened for binders within large chemical libraries. Screens will be carried out in academic or industry hubs, selected for having track records in high-quality data generation. The distribution of proteins to and among the screening hubs will be centrally coordinated to avoid duplication of effort.

A key strategic decision was to select the data-generating modalities. Platform(s) that screen for direct binding of ligands to purified proteins would be implemented, for the following reasons:

  1. (1)

    Direct-binding assays eliminate the impractical requirement to develop bespoke functional assays for each protein, including the thousands of human proteins with no known activity. For proteins with known function, a functional assay might aid in the hit verification process, and in the further advancement of the chemical matter.

  2. (2)

    A single preparation of purified protein can be used both for the primary binding screen used for hit identification as well as for the secondary orthogonal biophysical assays24 used for hit verification.

  3. (3)

    Screening campaigns could begin immediately, using the many hundreds of human proteins that have been purified or can be readily purified in high quality and quantity, by the SGC, by industry, and by the wider protein and structural biology academic communities.

After considering many screening platforms, DEL25,26,27,28 and AS–MS29,30,31 were chosen. These two biophysical screening methods have been used successfully for a wide variety of proteins, have the potential to generate millions of high-quality data points per screen and have already demonstrated efficient hit-finding results in our hands for diverse proteins. In addition, data generated by these methods have a common experimental design and can be represented in a machine-readable format and aggregated into increasingly large datasets. The large size and high dimensionality of these data also leverage respective analytical techniques that have been extensively developed by the ML/AI community32,33 and employed in cheminformatics for drug discovery applications34.

DEL screening

DEL screening is an affinity-mediated technique that has been used for more than two decades as a tool for identifying compounds that bind proteins35,36,37,38. In this technology, pools of compounds, each covalently attached to an oligonucleotide whose sequence encodes its synthetic history (and therefore the presumed compound identity), are incubated with the protein. Proteins are then captured using an affinity tag and associated library members are separated from non-binders by washing. The DNA encoding the retained library members is then amplified and sequenced, allowing the synthetic history of each compound and their enrichment over the background to be determined. Historically, enriched library compounds were resynthesized off the DNA and tested for binding or activity in an orthogonal assay. The technology allows for probing an enormous chemical library (>1 trillion members), but it has limitations: the presence of DNA induces many false positives, synthesizing the many potential binders off-DNA is time consuming and costly, and the chemical diversity of the library members is restricted by the requirement to use reactions that are compatible with the presence of DNA.

Some of these limitations can be overcome by integrating ML/AI with DEL screening. In this iteration of DEL screening data analysis, the datasets, comprising billions of data points and including both positive and (critically) negative binding data, are used to train ML algorithms and build models to predict the molecular features of a binder26,28,39. These algorithms are then used to search for active molecules within the billions of commercially available compounds or compound collections internal to organizations. The compounds are then acquired and tested for binding to the purified proteins using orthogonal binding and/or functional assays. This strategy offers several potential advantages. First, it is faster and less expensive for most investigators to purchase molecules40 than to synthesize each of the enriched library compounds. Second, predictions are not restricted to the molecules in the DEL but can be made against the large, diverse, and more drug-like chemical space represented in pre-enumerated, synthetically accessible commercial libraries.

This conceptual DEL ML workflow was pioneered by McCloskey et al.26 using three precedented targets and the scalability and generalizability of the approach have been subsequently confirmed27,28,41,42 These encouraging results emboldened us to imagine a scaled-up process in which DEL screening datasets from hundreds to thousands of proteins, including detailed protocols and metadata, would be provided to the academic and industry communities without restriction in a standardized, ML-ready format43,44. By providing open access to these data (aircheck.ai), the ML/AI community would be enabled to make predictions that can be tested experimentally and to develop methods that can be benchmarked (Fig. 3). In the first datasets in AIRCHECK, the data include a 10:1 ratio of negative to positive training examples, and up to 1 million data points. Negative training examples were proportionally distributed to positive training examples on a per-library basis. These data have already been used to build models that have successfully predicted new micromolar binders for the WDR91 protein45.

Fig. 3: Schematic representation of DEL screen output data and ML/AI workflow.
figure 3

Affinity-mediated selection of DNA-encoded chemical library (DEL) members leads to the enrichment of potential binders. Deep sequencing is used to identify the DNA barcodes of enriched and unenriched DEL members. Output data are subsequently translated into chemical structures and their corresponding chemical fingerprints. Both positive (enriched) and negative (not enriched) DEL members are included in open machine learning (ML)-ready datasets. The datasets are used to train ML models that are in turn used to recognize and nominate potential small-molecule binders from ultralarge chemical libraries. These compounds are procured, and their binding is tested experimentally in biophysical and/or biochemical assays. All generated data (ML-ready datasets, including chemical structures and/or their corresponding fingerprints, ML models and ligand validation data) are made public in a purpose-built, cloud-based storage system called Artificial Intelligence-Ready CHEmiCal Knowledge base (AIRCHECK). AI, artificial intelligence.

Full size image

Initially, the DEL screens will be carried out in selected organizations that have a track record of success in applying ML to their DEL data43,44. Over time, any other company or academic that has robust DEL synthesis and screening infrastructure and that agrees to share relevant data openly and in a standardized, ML-ready format43,44 would be welcome to join the initiative.

AS–MS

AS–MS has emerged as a robust hit identification approach in the pharmaceutical industry46. In this method, pools of mass-differentiated compounds, typically up to 2,000, are first incubated with the protein. The protein and small molecules are then resolved chromatographically, and compounds that co-elute with the protein are subjected to liquid chromatography–mass spectrometry and unambiguously identified by their exact masses. Compound binding is then verified using an orthogonal functional or binding assay(s). The current upper limit of detection for compounds in most AS–MS platforms is an affinity constant in the 1–15 micromolar range46.

With some notable exceptions47,48,49,50, AS–MS has not been widely adopted as a small-molecule screening platform in academia, in part due to the significant infrastructure that is required, but mostly because cost-effective use of the infrastructure requires a pipeline of purified proteins in multi-milligram quantities. Given the ability to access thousands of purified proteins in these quantities in this project, AS–MS was prioritized as a screening platform (Fig. 4). To optimize screening capacity and throughput, we have elected to implement an off-line AS–MS method that screens affinity-tagged proteins (his, GFP or biotin) against pools of compounds, and then resolves the protein/compound complexes from the non-binding compounds by binding the tagged protein to the corresponding magnetic affinity microbeads50. This pipeline was piloted by screening a diverse set of 31 proteins against a small chemical library explicitly optimized for mass spectrometry screening, and binders were discovered for 11 proteins51.

Fig. 4: The AS–MS screening workflow.
figure 4

From left to right: protein affinity purification by pooling 500 compounds in each affinity-selection mass spectrometry (AS–MS) sample; liquid chromatography–mass spectrometry (LC–MS) analysis of AS–MS samples; automatic data processing to identify hits; validation of hits using orthogonal biophysical methods (for example, surface plasmon resonance); and uploading AS–MS data to the Artificial Intelligence-Ready CHEmiCal Knowledge base (AIRCHECK) database, which is freely accessible to the whole community.

Full size image

The primary binding data and metadata from both DEL and AS–MS screens, as well as the results from secondary biophysical assays are now being placed into AIRCHECK without restriction on use. Raw mass spectrometry data will also be made available via Metabolomics Workbench (https://www.metabolomicsworkbench.org/) or a similar vehicle.

Annotation and verification of screening data

With a priority to generate screening datasets for ML/AI applications, particular attention will be paid to data quality, data annotation and data availability — using learnings from the experiences of our industry partners and other public initiatives52. Data quality standards will be made openly available and implemented at three key levels: for the protein samples, for the DEL and AS–MS screening outputs, and for the hit annotation.

Proteins

Proteins entering screens must meet established experimental quality criteria and must also be accompanied by key metadata that might influence data interpretation and model building, such as purification conditions or the presence of metal ions.

Screening datasets

Primary AS–MS and DEL screening-derived datasets will be assessed for technical quality against a set of relevant parameters (Supplementary Fig. 3). For public DEL and AS–MS screens that pass quality checks, all the raw screening data will be placed into the public domain.

Secondary annotation of primary screening data

Both experimental screening platforms will generate false positive and false negative hits, and to maintain the quality of the datasets for ML/AI applications, true and false positives must be distinguished using orthogonal assays53. This is technically challenging because weaker binding compounds are often insoluble at the concentrations used for many biophysical or functional assays54,55, which readily leads to artefacts in any single assay. As a result, many candidate binders may have to be tested in several different assays to gain sufficient confidence in their veracity.

Given the technical challenges in analysing weakly binding compounds, it will be critical to agree on how much effort the project should invest in determining if a screening hit is a true binder and to communicate the limitations of each of the assays54,55 and the resulting data to the modelling community. The strategic decision is how to balance annotating the largest number of true positives in the dataset, which is optimal for model building and also provides practical and valuable insight into the ligandability of a protein, with investing considerable resources in characterizing weakly binding compounds, which reduces the number of proteins that can be screened. The CACHE competition has created a document that explains how to interpret the biophysical binding assays and how to identify potential artefacts56. Constant and close discussion between experimentalists and data scientists in the project will minimize misinterpretation or over-interpretation of the screening and hit-characterization data.

For this project, a generous threshold in nominating hits from the initial screen would be implemented. A target affinity threshold of 10 µM (KD value) would be set for the orthogonal assay, potentially with some target-specific leeway20. Ideally, all candidate hits arising from the first orthogonal assay would be tested in an additional assay. The outcome will be a robust and inclusive list of well-annotated positive binders with a KD of ≤10 µM.

Data consistency

To prioritize data consistency, secondary screening and data annotation will be centralized in well-equipped and experienced academic or commercial laboratories that follow standard operating procedures. Samples will be exchanged regularly among laboratories and tested to monitor and eliminate any inter-laboratory variability. These laboratories will have access to a range of orthogonal assay formats including some form of surface binding assay such as surface plasmon resonance57 or grating-coupled interferometry58, and other biophysical methods with reasonable throughput, such as spectral shift, microscale thermophoresis, NMR or thermal shift methodologies59,60,61. One of the complexities of the project is that for many of the novel proteins that will be screened, orthogonal assays will have to be built without the benefit of a positive control binder. If functional assays that confirm target modulation are readily available, they would add another layer of verification to the hit-confirmation process and provide invaluable insight into how to develop the ligand into a chemical probe.

Data management and access

To fully realize the value of the annotated protein–ligand datasets, data management approaches will be treated with equal diligence as the experimental methods. Accordingly, the project will adhere to the data management roadmap recently described by Edfeldt and colleagues23. This will include establishing a controlled vocabulary for experimental data, using automation and electronic laboratory notebooks whenever possible, centralizing the database architecture to facilitate data integration and providing comprehensive documentation. Raw data will be provided whenever possible and transparent and reproducible data processing will be performed, including choosing the most relevant data representation, defining the right training and test sets, and providing estimates of prediction uncertainty. The comprehensive data management plan and its attributes are outlined in Table 1.

Table 1 Data management features
Full size table

Benchmarking with experimental feedback

The intention of providing large, consistent and high-quality datasets to the community is to enable the development of computational and ML/AI hit-finding and hit optimization methods. The models will be focused in the near term on predicting binders and optimization strategies for proteins in the screening set and in the longer term to build foundation models of hit discovery and optimization.

To accelerate the development of these methods, the project will partner with organizations, including CASP, DREAM62 and CACHE15, that launch benchmarking challenges, including those in which predictions from the community will be tested experimentally and compared. Data used as input to challenges would be kept confidential while challenges are in progress, and a regular cadence of challenges and data release would be established. The value of benchmarking initiatives in computational biology was clearly established by CASP, which, for over 30 years, has driven and monitored progressive improvements in computational methods63,64.

Some of the proposed initial benchmarking challenges are listed in Table 2. As the project advances, other types of benchmarking challenges would probably be incorporated, including those that involve combining data from multiple platforms, not only from AS–MS and DEL screening but also from novel hit-finding screening platforms that may arise in the future. A combination of challenges that better represent a typical drug discovery screening pipeline may also have added value, including those that integrate some form of experimental or computational protein structural information. However, as even relatively simple challenges require significant logistics and the associated experimental costs are high, running more elaborate pipelines at the start of the project is probably too ambitious.

Table 2 Sample benchmarking challenges
Full size table

Participants will be encouraged to make their models open source and freely available to anyone for use directly from AIRCHECK. To encourage this, the costs of procuring compounds and testing them experimentally, partly or in full, would ideally be defrayed for qualified participants who make their ML/AI models publicly available and with permissive licenses.

From pilots to implementation

Pilot projects have laid foundational elements for this project. The capacities are now in place to

(1) produce more than 2,000 high-quality human proteins, most ‘never previously liganded’; (2) screen these proteins against compound libraries using AS–MS and DEL; (3) store and disseminate project data, with a robust data management plan and database architecture; (4) annotate screening data and test predictions; and (5) solicit community contributions and participation.

In the first year of the project, the individual elements will be scaled and integrated to create a data generation plan that balances the shorter-term goal of identifying hits for high-priority proteins with the longer-term goal of generating data that will advance computational hit finding. The most likely screening cascade will involve screening each protein first by AS–MS against an exploratory (~15k) library whose composition will be made openly available. The rationale is that this screen is scalable, yields a direct binding readout, is the most cost effective and will most rapidly identify those proteins that are readily ‘ligandable’. The exploratory screen will also flag proteins that have physiochemical properties that render them unsuitable for AS–MS or DEL and will not be screened further. For example, the exploratory screen would flag proteins that appear stable but that in fact have transiently unfolded regions that may bind large numbers of compounds nonspecifically.

Stable and monodisperse proteins that do not yield hits from the AS–MS exploratory screen, or for which greater chemical diversity or large datasets are required, will be channelled into screens with larger chemical libraries, using both AS–MS and DEL. The proposed screening cascade will be reviewed periodically and adjusted to optimize the process or incorporate other screening approaches as needed.

Encouraging community contributions

Active participation of the wider scientific community will be essential to meet the project goals. Robust community engagement will be made feasible only by adopting open science principles within the project. For clarity, this means compounds, data and algorithms developed using project resources will be made available without restriction on use, and without intellectual property constraints. This open science position provides a clarity of purpose and short-circuits what could be prolonged and complex discussions over ownership of compounds and algorithms. In keeping with this position, there will also be no restrictions on subsequent research or commercial use of data, chemical structures and algorithms generated using project resources. With this as background, community contributions in the following areas are envisioned (Table 3).

Table 3 Community contributions
Full size table

Protein scientists

Structural biologists, and protein scientists more broadly, often have unique expertise in purifying proteins in their scientific areas of interest. Community members would be encouraged to contribute their purified proteins to the screening process. For the project, this will expand the diversity of the protein–ligand datasets. For the contributing scientist, this could provide open access to hits that they can pursue without restriction in their own laboratories. Already >30 protein scientists have sent proteins to Toronto for AS–MS screening, including from Brazil, the UK, Canada, Germany, Sweden and the USA, and binders for 8 of these community proteins have already been identified, verified by surface plasmon resonance, and shared with the contributor (for example, Wang et al.51). Tapping into this diverse community at a larger scale will bring enormous scientific benefit, but will also add logistical burdens, so the project will need to implement this process carefully.

Data generation

Project screening data would be generated initially using the AS–MS and DEL screening platforms in selected hubs. However, there are clear advantages to expanding the number of participating screening laboratories, and the range of data generation technologies. Accordingly, new screening methodologies would be explored on a continual basis. To manage this process, a set of ~25 well-characterized, diverse and ligandable proteins that will have been screened comprehensively through all the initial platforms will serve as a technology test set for new screening hubs or technologies. The project board and its scientific advisers will review all data and provide recommendations about adding new centres or technologies.

Engaging computational scientists worldwide

Each screen will generate multiple GB-scale datasets, which may need to be downloaded and manipulated. The use of cloud resources will ensure the scalability of the AIRCHECK platform while allowing users to easily access the data and the computational resources for ML/AI modelling. It also allows users to leverage education or research credits from large cloud providers to support more equitable, diverse and inclusive access (for example, Google Cloud program for higher education in Africa65). Scientists from resource-poor environments will be actively encouraged to participate. We will also facilitate the development of open-source algorithms by collaborating closely with a project-associated global network of computational scientists, called MAINFRAME66.

Chemists

The synthetic and medicinal chemistry communities will be encouraged (for example, through the SGC’s Open Chemistry Networks) to design and/or generate molecules related to the hits to improve the original binders. Testing these compounds within the project may generate preliminary structure–activity relationships and provide confidence that the binder can be advanced. Chemists will also be encouraged to contribute compounds that are theoretically accessible through their chemistries to the emerging virtual screening library of all compounds that are synthetically accessible67.

Training and networking

The project will be generating data explicitly to promote the development of ML/AI algorithms and as such will be operating at the intersection of experimental, data and computational sciences. This will provide an excellent training environment for scientists seeking a working and operational knowledge of the various domains, and programmes for trainees will be established. Regular project meetings that prioritize scientific exchange between the various communities will be established.

Project structure and governance

The project will be structured as a pre-competitive, open science partnership in which compound assay data generated with project resources, including chemical structures of confirmed hits and algorithms, will be made available to the public under a license that requires attribution but that places no restriction on subsequent use. As stated previously, the rationale is pragmatic and evidence-based: pragmatic, in that it would be almost impossible to imagine a seamless cross-sectoral, cross-disciplinary and multinational collaboration that could operate under an agreement that allowed for the protection of potential intellectual property; and evidence- based, in that the development of ML/AI algorithms, in whatever field, advances most rapidly when provided with open data and with a mechanism to benchmark progress transparently68.

The project needs to involve scientists from both public and private sectors to access the wide range of skill sets and expertise that will be required. It will also involve funding from both public and private sectors to achieve the requisite scale (Fig. 5). The major funders from the public and private sectors will form a governing board that oversees all project activities, including financial, scientific and management. The governing board will also oversee risk management, including any potential security risks associated with the data and the algorithms developed in the project. The governance board will be mandated to balance the needs of private sector funders with those of the public sector and its funding bodies, and also to provide a fair and time-limited mechanism for project or community contributors to pursue selected scientific questions. The governance structure that is currently used by the SGC is suitable because it has been used successfully to govern mission-oriented public–private partnerships of this complexity and scale69.

Fig. 5: Project governance.
figure 5

The project is designed as a public–private partnership. The governance structure is designed to ensure efficient operation, strategic alignment and excellence in research. It integrates inputs from both public and private sector funders to direct a multitiered management system composed of specialized committees.

Full size image

A range of outcomes

The long-term aim of this project is to develop efficient computational hit-finding algorithms that can be used to generate freely available, small-molecule binders initially for thousands of proteins, and eventually for all relevant human proteins. However, over the course of the project, intermediate outcomes of considerable value will be generated, and these outcomes should be used as metrics to track and manage the project. Some of the key metrics are listed in Table 4.

Table 4 Metrics
Full size table

A range of benefits to all participants

Open-access public–private partnerships are structures to carry out projects that require skills distributed among a wide range of academic and industry scientists, that tackle problems that span the boundary of public and private interests, and that might otherwise be crippled by intellectual property negotiations. However, in return for ceding their potential intellectual property rights to the public good, funders and participants must feel that they gain more than they lose, directly or indirectly. Table 5 lists some of the benefits that this project will generate for participants.

Table 5 Benefits to project participants
Full size table